diffusers 0.23.1__py3-none-any.whl → 0.25.0__py3-none-any.whl

diffusers/models/transformer_2d.py

@@ -20,9 +20,9 @@ from torch import nn
 
 from ..configuration_utils import ConfigMixin, register_to_config
 from ..models.embeddings import ImagePositionalEmbeddings
-from ..utils import USE_PEFT_BACKEND, BaseOutput, deprecate
+from ..utils import USE_PEFT_BACKEND, BaseOutput, deprecate, is_torch_version
 from .attention import BasicTransformerBlock
-from .embeddings import CaptionProjection, PatchEmbed
+from .embeddings import PatchEmbed, PixArtAlphaTextProjection
 from .lora import LoRACompatibleConv, LoRACompatibleLinear
 from .modeling_utils import ModelMixin
 from .normalization import AdaLayerNormSingle
@@ -70,6 +70,8 @@ class Transformer2DModel(ModelMixin, ConfigMixin):
             Configure if the `TransformerBlocks` attention should contain a bias parameter.
     """
 
+    _supports_gradient_checkpointing = True
+
     @register_to_config
     def __init__(
         self,
@@ -233,10 +235,14 @@ class Transformer2DModel(ModelMixin, ConfigMixin):
 
         self.caption_projection = None
         if caption_channels is not None:
-            self.caption_projection = CaptionProjection(in_features=caption_channels, hidden_size=inner_dim)
+            self.caption_projection = PixArtAlphaTextProjection(in_features=caption_channels, hidden_size=inner_dim)
 
         self.gradient_checkpointing = False
 
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
     def forward(
         self,
         hidden_states: torch.Tensor,
@@ -360,8 +366,19 @@ class Transformer2DModel(ModelMixin, ConfigMixin):
 
         for block in self.transformer_blocks:
             if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
                 hidden_states = torch.utils.checkpoint.checkpoint(
-                    block,
+                    create_custom_forward(block),
                     hidden_states,
                     attention_mask,
                     encoder_hidden_states,
@@ -369,7 +386,7 @@ class Transformer2DModel(ModelMixin, ConfigMixin):
                     timestep,
                     cross_attention_kwargs,
                     class_labels,
-                    use_reentrant=False,
+                    **ckpt_kwargs,
                 )
             else:
                 hidden_states = block(

diffusers/models/transformer_temporal.py

@@ -19,8 +19,10 @@ from torch import nn
 
 from ..configuration_utils import ConfigMixin, register_to_config
 from ..utils import BaseOutput
-from .attention import BasicTransformerBlock
+from .attention import BasicTransformerBlock, TemporalBasicTransformerBlock
+from .embeddings import TimestepEmbedding, Timesteps
 from .modeling_utils import ModelMixin
+from .resnet import AlphaBlender
 
 
 @dataclass
@@ -195,3 +197,183 @@ class TransformerTemporalModel(ModelMixin, ConfigMixin):
             return (output,)
 
         return TransformerTemporalModelOutput(sample=output)
+
+
+class TransformerSpatioTemporalModel(nn.Module):
+    """
+    A Transformer model for video-like data.
+
+    Parameters:
+        num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
+        in_channels (`int`, *optional*):
+            The number of channels in the input and output (specify if the input is **continuous**).
+        out_channels (`int`, *optional*):
+            The number of channels in the output (specify if the input is **continuous**).
+        num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
+        cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
+    """
+
+    def __init__(
+        self,
+        num_attention_heads: int = 16,
+        attention_head_dim: int = 88,
+        in_channels: int = 320,
+        out_channels: Optional[int] = None,
+        num_layers: int = 1,
+        cross_attention_dim: Optional[int] = None,
+    ):
+        super().__init__()
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+
+        inner_dim = num_attention_heads * attention_head_dim
+        self.inner_dim = inner_dim
+
+        # 2. Define input layers
+        self.in_channels = in_channels
+        self.norm = torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6)
+        self.proj_in = nn.Linear(in_channels, inner_dim)
+
+        # 3. Define transformers blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    inner_dim,
+                    num_attention_heads,
+                    attention_head_dim,
+                    cross_attention_dim=cross_attention_dim,
+                )
+                for d in range(num_layers)
+            ]
+        )
+
+        time_mix_inner_dim = inner_dim
+        self.temporal_transformer_blocks = nn.ModuleList(
+            [
+                TemporalBasicTransformerBlock(
+                    inner_dim,
+                    time_mix_inner_dim,
+                    num_attention_heads,
+                    attention_head_dim,
+                    cross_attention_dim=cross_attention_dim,
+                )
+                for _ in range(num_layers)
+            ]
+        )
+
+        time_embed_dim = in_channels * 4
+        self.time_pos_embed = TimestepEmbedding(in_channels, time_embed_dim, out_dim=in_channels)
+        self.time_proj = Timesteps(in_channels, True, 0)
+        self.time_mixer = AlphaBlender(alpha=0.5, merge_strategy="learned_with_images")
+
+        # 4. Define output layers
+        self.out_channels = in_channels if out_channels is None else out_channels
+        # TODO: should use out_channels for continuous projections
+        self.proj_out = nn.Linear(inner_dim, in_channels)
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        image_only_indicator: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ):
+        """
+        Args:
+            hidden_states (`torch.FloatTensor` of shape `(batch size, channel, height, width)`):
+                Input hidden_states.
+            num_frames (`int`):
+                The number of frames to be processed per batch. This is used to reshape the hidden states.
+            encoder_hidden_states ( `torch.LongTensor` of shape `(batch size, encoder_hidden_states dim)`, *optional*):
+                Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
+                self-attention.
+            image_only_indicator (`torch.LongTensor` of shape `(batch size, num_frames)`, *optional*):
+                A tensor indicating whether the input contains only images. 1 indicates that the input contains only
+                images, 0 indicates that the input contains video frames.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.transformer_temporal.TransformerTemporalModelOutput`] instead of a plain
+                tuple.
+
+        Returns:
+            [`~models.transformer_temporal.TransformerTemporalModelOutput`] or `tuple`:
+                If `return_dict` is True, an [`~models.transformer_temporal.TransformerTemporalModelOutput`] is
+                returned, otherwise a `tuple` where the first element is the sample tensor.
+        """
+        # 1. Input
+        batch_frames, _, height, width = hidden_states.shape
+        num_frames = image_only_indicator.shape[-1]
+        batch_size = batch_frames // num_frames
+
+        time_context = encoder_hidden_states
+        time_context_first_timestep = time_context[None, :].reshape(
+            batch_size, num_frames, -1, time_context.shape[-1]
+        )[:, 0]
+        time_context = time_context_first_timestep[None, :].broadcast_to(
+            height * width, batch_size, 1, time_context.shape[-1]
+        )
+        time_context = time_context.reshape(height * width * batch_size, 1, time_context.shape[-1])
+
+        residual = hidden_states
+
+        hidden_states = self.norm(hidden_states)
+        inner_dim = hidden_states.shape[1]
+        hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch_frames, height * width, inner_dim)
+        hidden_states = self.proj_in(hidden_states)
+
+        num_frames_emb = torch.arange(num_frames, device=hidden_states.device)
+        num_frames_emb = num_frames_emb.repeat(batch_size, 1)
+        num_frames_emb = num_frames_emb.reshape(-1)
+        t_emb = self.time_proj(num_frames_emb)
+
+        # `Timesteps` does not contain any weights and will always return f32 tensors
+        # but time_embedding might actually be running in fp16. so we need to cast here.
+        # there might be better ways to encapsulate this.
+        t_emb = t_emb.to(dtype=hidden_states.dtype)
+
+        emb = self.time_pos_embed(t_emb)
+        emb = emb[:, None, :]
+
+        # 2. Blocks
+        for block, temporal_block in zip(self.transformer_blocks, self.temporal_transformer_blocks):
+            if self.training and self.gradient_checkpointing:
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    block,
+                    hidden_states,
+                    None,
+                    encoder_hidden_states,
+                    None,
+                    use_reentrant=False,
+                )
+            else:
+                hidden_states = block(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                )
+
+            hidden_states_mix = hidden_states
+            hidden_states_mix = hidden_states_mix + emb
+
+            hidden_states_mix = temporal_block(
+                hidden_states_mix,
+                num_frames=num_frames,
+                encoder_hidden_states=time_context,
+            )
+            hidden_states = self.time_mixer(
+                x_spatial=hidden_states,
+                x_temporal=hidden_states_mix,
+                image_only_indicator=image_only_indicator,
+            )
+
+        # 3. Output
+        hidden_states = self.proj_out(hidden_states)
+        hidden_states = hidden_states.reshape(batch_frames, height, width, inner_dim).permute(0, 3, 1, 2).contiguous()
+
+        output = hidden_states + residual
+
+        if not return_dict:
+            return (output,)
+
+        return TransformerTemporalModelOutput(sample=output)

diffusers/models/unet_2d_blocks_flax.py

@@ -45,6 +45,7 @@ class FlaxCrossAttnDownBlock2D(nn.Module):
         dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
             Parameters `dtype`
     """
+
     in_channels: int
     out_channels: int
     dropout: float = 0.0
@@ -125,6 +126,7 @@ class FlaxDownBlock2D(nn.Module):
         dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
             Parameters `dtype`
     """
+
     in_channels: int
     out_channels: int
     dropout: float = 0.0
@@ -190,6 +192,7 @@ class FlaxCrossAttnUpBlock2D(nn.Module):
         dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
             Parameters `dtype`
     """
+
     in_channels: int
     out_channels: int
     prev_output_channel: int
@@ -275,6 +278,7 @@ class FlaxUpBlock2D(nn.Module):
         dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
             Parameters `dtype`
     """
+
     in_channels: int
     out_channels: int
     prev_output_channel: int
@@ -339,6 +343,7 @@ class FlaxUNetMidBlock2DCrossAttn(nn.Module):
         dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
             Parameters `dtype`
     """
+
     in_channels: int
     dropout: float = 0.0
     num_layers: int = 1

diffusers/models/unet_2d_condition.py

@@ -25,6 +25,7 @@ from .activations import get_activation
 from .attention_processor import (
     ADDED_KV_ATTENTION_PROCESSORS,
     CROSS_ATTENTION_PROCESSORS,
+    Attention,
     AttentionProcessor,
     AttnAddedKVProcessor,
     AttnProcessor,
@@ -794,6 +795,42 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
                 if hasattr(upsample_block, k) or getattr(upsample_block, k, None) is not None:
                     setattr(upsample_block, k, None)
 
+    def fuse_qkv_projections(self):
+        """
+        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query,
+        key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+        """
+        self.original_attn_processors = None
+
+        for _, attn_processor in self.attn_processors.items():
+            if "Added" in str(attn_processor.__class__.__name__):
+                raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
+
+        self.original_attn_processors = self.attn_processors
+
+        for module in self.modules():
+            if isinstance(module, Attention):
+                module.fuse_projections(fuse=True)
+
+    def unfuse_qkv_projections(self):
+        """Disables the fused QKV projection if enabled.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        """
+        if self.original_attn_processors is not None:
+            self.set_attn_processor(self.original_attn_processors)
+
     def forward(
         self,
         sample: torch.FloatTensor,
@@ -1022,6 +1059,15 @@ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin)
                 )
             image_embeds = added_cond_kwargs.get("image_embeds")
             encoder_hidden_states = self.encoder_hid_proj(image_embeds)
+        elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "ip_image_proj":
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'ip_image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            image_embeds = self.encoder_hid_proj(image_embeds).to(encoder_hidden_states.dtype)
+            encoder_hidden_states = torch.cat([encoder_hidden_states, image_embeds], dim=1)
+
         # 2. pre-process
         sample = self.conv_in(sample)
 

diffusers/models/unet_2d_condition_flax.py

@@ -100,18 +100,18 @@ class FlaxUNet2DConditionModel(nn.Module, FlaxModelMixin, ConfigMixin):
     sample_size: int = 32
     in_channels: int = 4
     out_channels: int = 4
-    down_block_types: Tuple[str] = (
+    down_block_types: Tuple[str, ...] = (
         "CrossAttnDownBlock2D",
         "CrossAttnDownBlock2D",
         "CrossAttnDownBlock2D",
         "DownBlock2D",
     )
-    up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")
+    up_block_types: Tuple[str, ...] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")
     only_cross_attention: Union[bool, Tuple[bool]] = False
-    block_out_channels: Tuple[int] = (320, 640, 1280, 1280)
+    block_out_channels: Tuple[int, ...] = (320, 640, 1280, 1280)
     layers_per_block: int = 2
-    attention_head_dim: Union[int, Tuple[int]] = 8
-    num_attention_heads: Optional[Union[int, Tuple[int]]] = None
+    attention_head_dim: Union[int, Tuple[int, ...]] = 8
+    num_attention_heads: Optional[Union[int, Tuple[int, ...]]] = None
     cross_attention_dim: int = 1280
     dropout: float = 0.0
     use_linear_projection: bool = False
@@ -120,7 +120,7 @@ class FlaxUNet2DConditionModel(nn.Module, FlaxModelMixin, ConfigMixin):
     freq_shift: int = 0
     use_memory_efficient_attention: bool = False
     split_head_dim: bool = False
-    transformer_layers_per_block: Union[int, Tuple[int]] = 1
+    transformer_layers_per_block: Union[int, Tuple[int, ...]] = 1
     addition_embed_type: Optional[str] = None
     addition_time_embed_dim: Optional[int] = None
     addition_embed_type_num_heads: int = 64
@@ -158,7 +158,7 @@ class FlaxUNet2DConditionModel(nn.Module, FlaxModelMixin, ConfigMixin):
             }
         return self.init(rngs, sample, timesteps, encoder_hidden_states, added_cond_kwargs)["params"]
 
-    def setup(self):
+    def setup(self) -> None:
         block_out_channels = self.block_out_channels
         time_embed_dim = block_out_channels[0] * 4
 
@@ -320,15 +320,15 @@ class FlaxUNet2DConditionModel(nn.Module, FlaxModelMixin, ConfigMixin):
 
     def __call__(
         self,
-        sample,
-        timesteps,
-        encoder_hidden_states,
+        sample: jnp.ndarray,
+        timesteps: Union[jnp.ndarray, float, int],
+        encoder_hidden_states: jnp.ndarray,
         added_cond_kwargs: Optional[Union[Dict, FrozenDict]] = None,
-        down_block_additional_residuals=None,
-        mid_block_additional_residual=None,
+        down_block_additional_residuals: Optional[Tuple[jnp.ndarray, ...]] = None,
+        mid_block_additional_residual: Optional[jnp.ndarray] = None,
         return_dict: bool = True,
         train: bool = False,
-    ) -> Union[FlaxUNet2DConditionOutput, Tuple]:
+    ) -> Union[FlaxUNet2DConditionOutput, Tuple[jnp.ndarray]]:
         r"""
         Args:
             sample (`jnp.ndarray`): (batch, channel, height, width) noisy inputs tensor